Xiaoyu Ma

Tencent Hunyuan Team, Ao Liu, Botong Zhou et al. · tencent-ai

As Large Language Models (LLMs) rapidly advance, we introduce Hunyuan-TurboS, a novel large hybrid Transformer-Mamba Mixture of Experts (MoE) model. It synergistically combines Mamba's long-sequence processing efficiency with Transformer's superior contextual understanding. Hunyuan-TurboS features an adaptive long-short chain-of-thought (CoT) mechanism, dynamically switching between rapid responses for simple queries and deep "thinking" modes for complex problems, optimizing computational resources. Architecturally, this 56B activated (560B total) parameter model employs 128 layers (Mamba2, Attention, FFN) with an innovative AMF/MF block pattern. Faster Mamba2 ensures linear complexity, Grouped-Query Attention minimizes KV cache, and FFNs use an MoE structure. Pre-trained on 16T high-quality tokens, it supports a 256K context length and is the first industry-deployed large-scale Mamba model. Our comprehensive post-training strategy enhances capabilities via Supervised Fine-Tuning (3M instructions), a novel Adaptive Long-short CoT Fusion method, Multi-round Deliberation Learning for iterative improvement, and a two-stage Large-scale Reinforcement Learning process targeting STEM and general instruction-following. Evaluations show strong performance: overall top 7 rank on LMSYS Chatbot Arena with a score of 1356, outperforming leading models like Gemini-2.0-Flash-001 (1352) and o4-mini-2025-04-16 (1345). TurboS also achieves an average of 77.9% across 23 automated benchmarks. Hunyuan-TurboS balances high performance and efficiency, offering substantial capabilities at lower inference costs than many reasoning models, establishing a new paradigm for efficient large-scale pre-trained models.

Likun Zhang, Xiaoyu Ma, Christopher K. Wikle et al.

Many real-world processes have complex tail dependence structures that cannot be characterized using classical Gaussian processes. More flexible spatial extremes models exhibit appealing extremal dependence properties but are often exceedingly prohibitive to fit and simulate from in high dimensions. In this paper, we aim to push the boundaries on computation and modeling of high-dimensional spatial extremes via integrating a new spatial extremes model that has flexible and non-stationary dependence properties in the encoding-decoding structure of a variational autoencoder called the XVAE. The XVAE can emulate spatial observations and produce outputs that have the same statistical properties as the inputs, especially in the tail. Our approach also provides a novel way of making fast inference with complex extreme-value processes. Through extensive simulation studies, we show that our XVAE is substantially more time-efficient than traditional Bayesian inference while outperforming many spatial extremes models with a stationary dependence structure. Lastly, we analyze a high-resolution satellite-derived dataset of sea surface temperature in the Red Sea, which includes 30 years of daily measurements at 16703 grid cells. We demonstrate how to use XVAE to identify regions susceptible to marine heatwaves under climate change and examine the spatial and temporal variability of the extremal dependence structure.

Li Kang, Heng Zhou, Xiufeng Song et al.

Recent advancements in multimodal large language models and vision-languageaction models have significantly driven progress in Embodied AI. As the field transitions toward more complex task scenarios, multi-agent system frameworks are becoming essential for achieving scalable, efficient, and collaborative solutions. This shift is fueled by three primary factors: increasing agent capabilities, enhancing system efficiency through task delegation, and enabling advanced human-agent interactions. To address the challenges posed by multi-agent collaboration, we propose the Multi-Agent Robotic System (MARS) Challenge, held at the NeurIPS 2025 Workshop on SpaVLE. The competition focuses on two critical areas: planning and control, where participants explore multi-agent embodied planning using vision-language models (VLMs) to coordinate tasks and policy execution to perform robotic manipulation in dynamic environments. By evaluating solutions submitted by participants, the challenge provides valuable insights into the design and coordination of embodied multi-agent systems, contributing to the future development of advanced collaborative AI systems.

Xiaoyu Ma, Weijie Zhang, Yuanhao Gao et al.

Multimodal learning often suffers from modality imbalance, where modalities that converge faster dominate optimization while others remain undertrained. Existing approaches typically mitigate this issue by strengthening the weak modality or adjusting optimization gradients. However, such strategies mainly compensate for optimization rate discrepancies, often at the expense of the strong modality's optimization capacity, without analyzing how these discrepancies arise at the modality level. Based on theoretical insights and empirical observations, we argue that the discrepancy of learning pace arises from differences in the mapping difficulty between modality-specific feature space and the shared label space. To address this issue, we propose Balanced Multimodal Label Reshaping (BMLR), the first method that promotes multimodal balance from the label-side design. BMLR reshapes the cross-modal label space to equalize mapping difficulty across modalities, thereby facilitating modality interaction and injecting richer inter-class information into each modality. Extensive experiments across multiple architectures demonstrate that BMLR consistently improves multimodal performance and exhibits strong compatibility with diverse model designs. The source code will be released soon.

Yenho Chen, Carl W. Harris, Xiaoyu Ma et al.

We propose a decoding-based approach to detect context effects on neural codes in longitudinal neural recording data. The approach is agnostic to how information is encoded in neural activity, and can control for a variety of possible confounding factors present in the data. We demonstrate our approach by determining whether it is possible to decode location encoding from prefrontal cortex in the mouse and, further, testing whether the encoding changes due to task engagement.

Xiaoyu Ma, Yiwen Li, Haoyue Liu et al.

Automatic prompt optimization (APO) hinges on the quality of its evaluation signal, yet scoring every prompt candidate on the full training set is prohibitively expensive. Existing methods either fix a single evaluation subset before optimization begins (principled but prompt-agnostic) or adapt it heuristically during optimization (flexible but unstable and lacking formal guarantees). We observe that APO naturally maps to an online adaptive testing problem: prompts are examinees, training examples are test items, and the scheduler should select items that best discriminate among the strongest candidates. This insight motivates Prompt-Aware Online Evaluation Scheduling (POES), which integrates an IRT-based discrimination utility, a facility-location coverage term, and switching-cost-aware warm-start swaps into a unified objective that is provably monotone submodular, yielding a (1-1/e) greedy guarantee for cold starts and bounded drift for warm-start updates. An adaptive controller modulates the exploration-exploitation balance based on optimization progress. Across 36 tasks spanning three benchmark families, POES achieves the highest overall average accuracy (6.2 percent improvement over the best baseline) with negligible token overhead (approximately 4 percent) at the same evaluation budget. Moreover, principled selection at k = 20 examples matches or exceeds the performance of naive evaluation at k = 30-50, reducing token consumption by 35-60 percent, showing that selecting smarter is more effective than selecting more. Our results demonstrate that evaluation scheduling is a first-class component of APO, not an implementation detail.

Xiaoyu Ma, Lianyu Hu, Wenbing Tang et al.

Embodied task planning requires agents to execute long-horizon, goal-directed actions in complex 3D environments, where success depends on both immediate perception and accumulated experience across tasks. However, most existing LLM-based planners are stateless and reactive, operating without persistent memory and therefore repeating errors and struggling with spatial or temporal dependencies. We propose BrainMem(Brain-Inspired Evolving Memory), a training-free hierarchical memory system that equips embodied agents with working, episodic, and semantic memory inspired by human cognition. BrainMem continuously transforms interaction histories into structured knowledge graphs and distilled symbolic guidelines, enabling planners to retrieve, reason over, and adapt behaviors from past experience without any model fine-tuning or additional training. This plug-and-play design integrates seamlessly with arbitrary multi-modal LLMs and greatly reduces reliance on task-specific prompt engineering. Extensive experiments on four representative benchmarks, including EB-ALFRED, EB-Navigation, EB-Manipulation, and EB-Habitat, demonstrate that BrainMem significantly enhances task success rates across diverse models and difficulty subsets, with the largest gains observed on long-horizon and spatially complex tasks. These results highlight evolving memory as a promising and scalable mechanism for generalizable embodied intelligence.

Xiaoyu Ma, Fang Fang, Ximing Xie et al.

The limited number of qubits is a major bottleneck in Quantum Approximate Optimization Algorithm (QAOA) for large-scale combinatorial optimization in the Noisy Intermediate-Scale Quantum (NISQ) era. To make progress, existing techniques rely on qubit reduction at the cost of information loss, hence leading to degraded computational performance. As a remedy, we propose the Equivalence-preserving Qubit Efficient QAOA (EQE-QAOA), which significantly reduces the required number of qubits without degrading the performance of QAOA. By exploiting intrinsic symmetries and conserved quantities, we first demonstrate that the QAOA dynamics are strictly confined to an invariant subspace of the Hilbert space. We subsequently prove that the evolution within this subspace is exactly equivalent to that of the full-scale system, achieving the same optimal solution as the original QAOA. Moreover, to reduce the number of qubits, we propose an isometric mapping that re-encodes the subspace into a space relying on fewer qubits. Furthermore, we derive the applicability conditions of EQE-QAOA and show that it is broadly applicable to large-scale combinatorial optimization problems, excluding only unconstrained problems with completely independent variables. Numerical simulations based on Max-Cut instances validate that EQE-QAOA significantly reduces qubit requirements and computational resources, while preserving exact optimization performance.

Xiaoyu Ma, Hao Chen, Yongjian Deng

Different modalities hold considerable gaps in optimization trajectories, including speeds and paths, which lead to modality laziness and modality clash when jointly training multimodal models, resulting in insufficient and imbalanced multimodal learning. Existing methods focus on enforcing the weak modality by adding modality-specific optimization objectives, aligning their optimization speeds, or decomposing multimodal learning to enhance unimodal learning. These methods fail to achieve both unimodal sufficiency and multimodal balance. In this paper, we, for the first time, address both concerns by proposing multimodal Data Remixing, including decoupling multimodal data and filtering hard samples for each modality to mitigate modality imbalance; and then batch-level reassembling to align the gradient directions and avoid cross-modal interference, thus enhancing unimodal learning sufficiency. Experimental results demonstrate that our method can be seamlessly integrated with existing approaches, improving accuracy by approximately 6.50%$\uparrow$ on CREMAD and 3.41%$\uparrow$ on Kinetic-Sounds, without training set expansion or additional computational overhead during inference. The source code is available at https://github.com/MatthewMaxy/Remix_ICML2025.

Xiaoyu Ma, David Patterson

Large Language Model (LLM) inference is hard. The autoregressive Decode phase of the underlying Transformer model makes LLM inference fundamentally different from training. Exacerbated by recent AI trends, the primary challenges are memory and interconnect rather than compute. To address these challenges, we highlight four architecture research opportunities: High Bandwidth Flash for 10X memory capacity with HBM-like bandwidth; Processing-Near-Memory and 3D memory-logic stacking for high memory bandwidth; and low-latency interconnect to speedup communication. While our focus is datacenter AI, we also review their applicability for mobile devices.

Sylvain Sardy, Maxime van Cutsem, Xiaoyu Ma

The growing environmental footprint of artificial intelligence (AI), especially in terms of storage and computation, calls for more frugal and interpretable models. Sparse models (e.g., linear, neural networks) offer a promising solution by selecting only the most relevant features, reducing complexity, preventing over-fitting and enabling interpretation-marking a step towards truly intelligent AI. The concept of a right amount of sparsity (without too many false positive or too few true positive) is subjective. So we propose a new paradigm previously only observed and mathematically studied for compressed sensing (noiseless linear models): obtaining a phase transition in the probability of retrieving the relevant features. We show in practice how to obtain this phase transition for a class of sparse learners. Our approach is flexible and applicable to complex models ranging from linear to shallow and deep artificial neural networks while supporting various loss functions and sparsity-promoting penalties. It does not rely on cross-validation or on a validation set to select its single regularization parameter. For real-world data, it provides a good balance between predictive accuracy and feature sparsity. A Python package is available at https://github.com/VcMaxouuu/HarderLASSO containing all the simulations and ready-to-use models.

Jiayi Wang, Yi-An Mao, Xiaoyu Ma et al.

Optic nerve head (ONH) detection has been a crucial area of study in ophthalmology for years. However, the significant discrepancy between fundus image datasets, each generated using a single type of fundus camera, poses challenges to the generalizability of ONH detection approaches developed based on semantic segmentation networks. Despite the numerous recent advancements in general-purpose semantic segmentation methods using convolutional neural networks (CNNs) and Transformers, there is currently a lack of benchmarks for these state-of-the-art (SoTA) networks specifically trained for ONH detection. Therefore, in this article, we make contributions from three key aspects: network design, the publication of a dataset, and the establishment of a comprehensive benchmark. Our newly developed ONH detection network, referred to as ODFormer, is based upon the Swin Transformer architecture and incorporates two novel components: a multi-scale context aggregator and a lightweight bidirectional feature recalibrator. Our published large-scale dataset, known as TongjiU-DROD, provides multi-resolution fundus images for each participant, captured using two distinct types of cameras. Our established benchmark involves three datasets: DRIONS-DB, DRISHTI-GS1, and TongjiU-DROD, created by researchers from different countries and containing fundus images captured from participants of diverse races and ages. Extensive experimental results demonstrate that our proposed ODFormer outperforms other state-of-the-art (SoTA) networks in terms of performance and generalizability. Our dataset and source code are publicly available at mias.group/ODFormer.

Jingxin Mao, Xiaoyu Ma, Yanlong Bi et al.

Diabetic retinopathy (DR), as a debilitating ocular complication, necessitates prompt intervention and treatment. Despite the effectiveness of artificial intelligence in aiding DR grading, the progression of research toward enhancing the interpretability of DR grading through precise lesion segmentation faces a severe hindrance due to the scarcity of pixel-level annotated DR datasets. To mitigate this, this paper presents and delineates TJDR, a high-quality DR pixel-level annotation dataset, which comprises 561 color fundus images sourced from the Tongji Hospital Affiliated to Tongji University. These images are captured using diverse fundus cameras including Topcon's TRC-50DX and Zeiss CLARUS 500, exhibit high resolution. For the sake of adhering strictly to principles of data privacy, the private information of images is meticulously removed while ensuring clarity in displaying anatomical structures such as the optic disc, retinal blood vessels, and macular fovea. The DR lesions are annotated using the Labelme tool, encompassing four prevalent DR lesions: Hard Exudates (EX), Hemorrhages (HE), Microaneurysms (MA), and Soft Exudates (SE), labeled respectively from 1 to 4, with 0 representing the background. Significantly, experienced ophthalmologists conduct the annotation work with rigorous quality assurance, culminating in the construction of this dataset. This dataset has been partitioned into training and testing sets and publicly released to contribute to advancements in the DR lesion segmentation research community.

Xiaoyu Ma, Hao Chen

Multimodal learning integrates information from different modalities to enhance model performance, yet it often suffers from modality imbalance, where dominant modalities overshadow weaker ones during joint optimization. This paper reveals that such an imbalance not only occurs during representation learning but also manifests significantly at the decision layer. Experiments on audio-visual datasets (CREMAD and Kinetic-Sounds) show that even after extensive pretraining and balanced optimization, models still exhibit systematic bias toward certain modalities, such as audio. Further analysis demonstrates that this bias originates from intrinsic disparities in feature-space and decision-weight distributions rather than from optimization dynamics alone. We argue that aggregating uncalibrated modality outputs at the fusion stage leads to biased decision-layer weighting, hindering weaker modalities from contributing effectively. To address this, we propose that future multimodal systems should focus more on incorporate adaptive weight allocation mechanisms at the decision layer, enabling relative balanced according to the capabilities of each modality.

Xiaoyu Ma, Ding Ding, Hao Chen

Multimodal video understanding plays a crucial role in tasks such as action recognition and emotion classification by combining information from different modalities. However, multimodal models are prone to overfitting strong modalities, which can dominate learning and suppress the contributions of weaker ones. To address this challenge, we first propose Multimodal Mixup (MM), which applies the Mixup strategy at the aggregated multimodal feature level to mitigate overfitting by generating virtual feature-label pairs. While MM effectively improves generalization, it treats all modalities uniformly and does not account for modality imbalance during training. Building on MM, we further introduce Balanced Multimodal Mixup (B-MM), which dynamically adjusts the mixing ratios for each modality based on their relative contributions to the learning objective. Extensive experiments on several datasets demonstrate the effectiveness of our methods in improving generalization and multimodal robustness.

Zijie Yang, Qiji Zhou, Fang Guo et al.

Research data are the foundation of Artificial Intelligence (AI)-driven science, yet current AI applications remain limited to a few fields with readily available, well-structured, digitized datasets. Achieving comprehensive AI empowerment across multiple disciplines is still out of reach. Present-day research data collection is often fragmented, lacking unified standards, inefficiently managed, and difficult to share. Creating a single platform for standardized data digitization needs to overcome the inherent challenge of balancing between universality (supporting the diverse, ever-evolving needs of various disciplines) and standardization (enforcing consistent formats to fully enable AI). No existing platform accommodates both facets. Building a truly multidisciplinary platform requires integrating scientific domain knowledge with sophisticated computing skills. Researchers often lack the computational expertise to design customized and standardized data recording methods, whereas platform developers rarely grasp the intricate needs of multiple scientific domains. These gaps impede research data standardization and hamper AI-driven progress. In this study, we address these challenges by developing Airalogy (https://airalogy.com), the world's first AI- and community-driven platform that balances universality and standardization for digitizing research data across multiple disciplines. Airalogy represents entire research workflows using customizable, standardized data records and offers an advanced AI research copilot for intelligent Q&A, automated data entry, analysis, and research automation. Already deployed in laboratories across all four schools of Westlake University, Airalogy has the potential to accelerate and automate scientific innovation in universities, industry, and the global research community-ultimately benefiting humanity as a whole.

Xiaoyu Ma, Sylvain Sardy, Nick Hengartner et al.

To fit sparse linear associations, a LASSO sparsity inducing penalty with a single hyperparameter provably allows to recover the important features (needles) with high probability in certain regimes even if the sample size is smaller than the dimension of the input vector (haystack). More recently learners known as artificial neural networks (ANN) have shown great successes in many machine learning tasks, in particular fitting nonlinear associations. Small learning rate, stochastic gradient descent algorithm and large training set help to cope with the explosion in the number of parameters present in deep neural networks. Yet few ANN learners have been developed and studied to find needles in nonlinear haystacks. Driven by a single hyperparameter, our ANN learner, like for sparse linear associations, exhibits a phase transition in the probability of retrieving the needles, which we do not observe with other ANN learners. To select our penalty parameter, we generalize the universal threshold of Donoho and Johnstone (1994) which is a better rule than the conservative (too many false detections) and expensive cross-validation. In the spirit of simulated annealing, we propose a warm-start sparsity inducing algorithm to solve the high-dimensional, non-convex and non-differentiable optimization problem. We perform precise Monte Carlo simulations to show the effectiveness of our approach.

Amirali Boroumand, Saugata Ghose, Berkin Akin et al.

Emerging edge computing platforms often contain machine learning (ML) accelerators that can accelerate inference for a wide range of neural network (NN) models. These models are designed to fit within the limited area and energy constraints of the edge computing platforms, each targeting various applications (e.g., face detection, speech recognition, translation, image captioning, video analytics). To understand how edge ML accelerators perform, we characterize the performance of a commercial Google Edge TPU, using 24 Google edge NN models (which span a wide range of NN model types) and analyzing each NN layer within each model. We find that the Edge TPU suffers from three major shortcomings: (1) it operates significantly below peak computational throughput, (2) it operates significantly below its theoretical energy efficiency, and (3) its memory system is a large energy and performance bottleneck. Our characterization reveals that the one-size-fits-all, monolithic design of the Edge TPU ignores the high degree of heterogeneity both across different NN models and across different NN layers within the same NN model, leading to the shortcomings we observe. We propose a new acceleration framework called Mensa. Mensa incorporates multiple heterogeneous edge ML accelerators (including both on-chip and near-data accelerators), each of which caters to the characteristics of a particular subset of NN models and layers. During NN inference, for each NN layer, Mensa decides which accelerator to schedule the layer on, taking into account both the optimality of each accelerator for the layer and layer-to-layer communication costs. Averaged across all 24 Google edge NN models, Mensa improves energy efficiency and throughput by 3.0x and 3.1x over the Edge TPU, and by 2.4x and 4.3x over Eyeriss~v2, a state-of-the-art accelerator.

Amirali Boroumand, Saugata Ghose, Berkin Akin et al.

As the need for edge computing grows, many modern consumer devices now contain edge machine learning (ML) accelerators that can compute a wide range of neural network (NN) models while still fitting within tight resource constraints. We analyze a commercial Edge TPU using 24 Google edge NN models (including CNNs, LSTMs, transducers, and RCNNs), and find that the accelerator suffers from three shortcomings, in terms of computational throughput, energy efficiency, and memory access handling. We comprehensively study the characteristics of each NN layer in all of the Google edge models, and find that these shortcomings arise from the one-size-fits-all approach of the accelerator, as there is a high amount of heterogeneity in key layer characteristics both across different models and across different layers in the same model. We propose a new acceleration framework called Mensa. Mensa incorporates multiple heterogeneous ML edge accelerators (including both on-chip and near-data accelerators), each of which caters to the characteristics of a particular subset of models. At runtime, Mensa schedules each layer to run on the best-suited accelerator, accounting for both efficiency and inter-layer dependencies. As we analyze the Google edge NN models, we discover that all of the layers naturally group into a small number of clusters, which allows us to design an efficient implementation of Mensa for these models with only three specialized accelerators. Averaged across all 24 Google edge models, Mensa improves energy efficiency and throughput by 3.0x and 3.1x over the Edge TPU, and by 2.4x and 4.3x over Eyeriss v2, a state-of-the-art accelerator.